Launch Unbind Flash V1 with source-specific evaluation graphics

.gitattributes

@@ -36,3 +36,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 DeepSeek-V4-Flash-Q2_K_S-00001-of-00002.gguf filter=lfs diff=lfs merge=lfs -text
 DeepSeek-V4-Flash-Q2_K_S-00002-of-00002.gguf filter=lfs diff=lfs merge=lfs -text
 assets/unbind-flash-v1.png filter=lfs diff=lfs merge=lfs -text
+assets/hosted-capability.png filter=lfs diff=lfs merge=lfs -text
+assets/hosted-refusal.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -11,31 +11,48 @@ tags:
 - conversational
 ---
 
-![Unbind Flash V1 — Your mind, unbound.](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/unbind-flash-v1.png)
+![Unbind Flash V1 — Go off script.](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/unbind-flash-v1.png)
 
 # Unbind Flash V1
 
-**Room to think. Space to explore.**
+**Go off script.**
 
-Unbind Flash V1 is an Unbind-branded distribution of DeepSeek-V4-Flash in GGUF format. It packages the Q2_K_S quantization for people who want to download and work with the model locally.
+Today we're introducing Unbind Flash V1: an uncensored hosted edition and a separate public GGUF download.
 
-| Model | Format | Quantization | Download size | License |
-|:--|:--|:--|:--|:--|
-| Unbind Flash V1 | GGUF | Q2_K_S | ~98.6 GB, two shards | MIT |
+Unbind is for the question you haven't quite figured out how to ask. The idea that sounds unreasonable. The argument worth taking apart. We want conversations with room to develop—and a product that invites you to keep going.
 
-## About this release
+We're excited to put Unbind into people's hands. Explore it, challenge it, and tell us where it needs work.
 
-This release uses the unchanged Q2_K_S weights published by [ggml-org](https://huggingface.co/ggml-org/DeepSeek-V4-Flash-GGUF), based on [DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash) by DeepSeek. UnbindAI added the branding and model card; it did not train or fine-tune the weights. The source notes that its Q2 quantizations do not use imatrix calibration.
+## Choose your edition
 
-Both GGUF shards are needed. Storage size does not include the additional runtime memory required to load and run the model. Compatibility depends on your inference software and hardware; consult the upstream model pages for details.
+| Edition | What it contains |
+|:--|:--|
+| **Hosted Unbind Flash V1** | A separate hosted experience using the [dealignai 1017 UNCENSORED FP8 checkpoint](https://huggingface.co/dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8), identified by UnbindAI. |
+| **This repository's download** | The unchanged [ggml-org DeepSeek-V4-Flash Q2_K_S GGUF](https://huggingface.co/ggml-org/DeepSeek-V4-Flash-GGUF): two shards, approximately **98.6 GB**, MIT licensed. This is an earlier model and is not the hosted abliterated checkpoint. |
 
-## Evaluation and limitations
+## Published results for the hosted checkpoint
 
-UnbindAI has not independently tested or evaluated this distribution. Model outputs can be inaccurate or unsuitable for a particular use. Review the upstream model's documentation and assess the model for your own use.
+These are the source publisher's measurements of the checkpoint identified for hosted Unbind. They are not Unbind endpoint tests or measurements of this repository's GGUF files.
 
-## License and credits
+![Publisher-reported MMLU comparison](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/hosted-capability.png)
 
-Released under the [MIT License](LICENSE). The license retains DeepSeek's copyright and permission notice. The model was developed by DeepSeek; the GGUF conversion and quantization were provided by ggml-org.
+The publisher reports **83.77% (955/1,140)** versus **87.63% (999/1,140)** for the base on a 57-subject MMLU sample. Prompts and decoding matched; serving stacks differed.
+
+![Publisher-reported HarmBench classifications](https://huggingface.co/UnbindAI/Unbind-Flash-V1/resolve/main/assets/hosted-refusal.png)
+
+The publisher reports **320/320 classified COMPLY**, including **0/320 explicit refusals**, with max reasoning and a 200-new-token cap. Its classifier can score the opening reasoning trace when no final answer is available. This result does not establish completed answers or universal compliance.
+
+[Read the source evaluation and method](https://huggingface.co/dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8#internal-benchmarks-greedy-t0-this-repo).
+
+## Download details
+
+Download both GGUF shards from **Files and versions**. Runtime memory requirements exceed the file size and depend on your software, hardware and context length. The source notes that this Q2 quantization does not use imatrix calibration.
+
+The downloadable weights were copied unchanged. UnbindAI created this distribution's branding and documentation and has not trained, fine-tuned or independently evaluated these weights. The hosted edition uses a different checkpoint.
+
+## Credits and license
+
+The downloadable base model was developed by DeepSeek; the GGUF conversion and quantization were published by ggml-org. The hosted checkpoint's weight-level modification was published by dealignai. Both sources identify their license as MIT; this repository retains the original [MIT license notice](https://huggingface.co/UnbindAI/Unbind-Flash-V1/blob/main/LICENSE).
 
 ---